A non-convex relaxed version of minimax theorems

M.I.A. Ghitri,A. Hantoute

2023-08-18

Abstract:Given a subset $A\times B$ of a locally convex space $X\times Y$ (with $A$ compact) and a function $f:A\times B\rightarrow\overline{\mathbb{R}}$ such that $f(\cdot,y),$ $y\in B,$ are concave and upper semicontinuous, the minimax inequality $\max_{x\in A} \inf_{y\in B} f(x,y) \geq \inf_{y\in B} \sup_{x\in A_{0}} f(x,y)$ is shown to hold provided that $A_{0}$ be the set of $x\in A$ such that $f(x,\cdot)$ is proper, convex and lower semi-contiuous. Moreover, if in addition $A\times B\subset f^{-1}(\mathbb{R})$, then we can take as $A_{0}$ the set of $x\in A$ such that $f(x,\cdot)$ is convex. The relation to Moreau's biconjugate representation theorem is discussed, and some applications to\ convex duality are provided. Key words. Minimax theorem, Moreau theorem, conjugate function, convex optimization.

Optimization and Control

What problem does this paper attempt to address?

This paper attempts to solve two main problems related to the classical minimax theorem:

1. **Can the finiteness assumption of the function \( f \) be removed?**
Specifically, can the minimax inequality still hold without assuming \( A\times B\subset f^{-1}(\mathbb{R}) \)? This would make the theorem applicable to a broader family of functions, especially convex functions that take infinite values at some points, which are very common in areas such as convex optimization.

2. **If the function \( f(x,\cdot) \) is not convex for all \( x\in A \), but only convex on a subset of \( A \), does the minimax inequality still hold?**
In other words, if we replace \( A \) with \( A_0:=\{x\in A: f(x,\cdot)\text{ is convex}\} \), does the minimax inequality still hold?

To answer these two questions, the author proposes two relaxed versions of the minimax theorem:

- **Theorem 7**: For a given function \( f:X\times Y\rightarrow\mathbb{R} \), if \( A_0 := \{x\in A: f(x,\cdot)\in\Gamma_0(Y)\} \), then
\[
\inf_{y\in Y}\sup_{x\in A_0}f(x,y)\leq\max_{x\in A}\inf_{y\in Y}f(x,y).
\]

- **Theorem 9**: If \( A\times B\subset f^{-1}(\mathbb{R}) \) and \( A_1 := \{x\in A: f(x,\cdot)\text{ is convex}\} \), then
\[
\inf_{y\in B}\sup_{x\in A_1}f(x,y)\leq\max_{x\in A}\inf_{y\in B}f(x,y).
\]

The proofs of these two theorems are based on Moreau's biconjugate representation theorem, and it is proved that the minimax theorem is equivalent to Moreau's theorem. These results have wide applications in optimization and convex duality theory.